diff --git a/.github/workflows/ci.yaml b/.github/workflows/ci.yaml
index 7711409..519d8f5 100644
--- a/.github/workflows/ci.yaml
+++ b/.github/workflows/ci.yaml
@@ -364,12 +364,18 @@ jobs:
           key: ${{ runner.os }}-go-${{ hashFiles('go.sum') }}
           restore-keys: ${{ runner.os }}-go-
 
+      # This cache holds analyzer facts (staticcheck's "does this call
+      # return?"), so a stale or partial entry does not just cost time -- it
+      # turns checks like SA5011 into false positives on untouched code. Hence
+      # the lock file in the key, so a golangci-lint or Go bump starts fresh
+      # rather than inheriting the previous binary's facts, and no restore-keys:
+      # a miss must start cold instead of falling back to the newest entry
+      # under the prefix, which is how one bad entry outlives its own commit.
       - name: Cache golangci-lint
         uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.cache/golangci-lint
-          key: ${{ runner.os }}-golangci-lint-${{ hashFiles('go.sum', '.golangci.yml') }}
-          restore-keys: ${{ runner.os }}-golangci-lint-
+          key: ${{ runner.os }}-golangci-lint-${{ hashFiles('go.sum', '.golangci.yml', '.mise/mise.lock') }}
 
       - name: Run linter
         run: mise run lint
@@ -516,12 +522,15 @@ jobs:
     runs-on: ubuntu-24.04
     permissions: {}
     steps:
+      # cancelled counts as failed: !cancelled() above still runs this job when
+      # individual needs were cancelled but the run was not, and a required job
+      # that never reached a verdict must not clear the merge gate.
       - name: Any jobs failed?
-        if: ${{ contains(needs.*.result, 'failure') }}
+        if: ${{ contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled') }}
         run: |
           exit 1
 
       - name: All jobs passed or skipped?
-        if: ${{ !(contains(needs.*.result, 'failure')) }}
+        if: ${{ !(contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled')) }}
         run: |
           echo "All jobs passed or skipped" && echo "${{ toJSON(needs.*.result) }}"
diff --git a/internal/agent/poolstats.go b/internal/agent/poolstats.go
index 7fa3542..d00c814 100644
--- a/internal/agent/poolstats.go
+++ b/internal/agent/poolstats.go
@@ -179,7 +179,8 @@ func (p *PoolStatsPublisher) publish(ctx context.Context) error {
 			return err
 		}
 		if high && !wasHigh && p.Recorder != nil {
-			p.Recorder.Eventf(cur, nil, corev1.EventTypeWarning, ConditionPoolUsageHigh, "Sample", msg)
+			// The message is data, not a format string: it embeds usage percentages.
+			p.Recorder.Eventf(cur, nil, corev1.EventTypeWarning, ConditionPoolUsageHigh, "Sample", "%s", msg)
 		}
 		return nil
 	})
diff --git a/internal/agent/poolstats_test.go b/internal/agent/poolstats_test.go
index 89ab158..60a4581 100644
--- a/internal/agent/poolstats_test.go
+++ b/internal/agent/poolstats_test.go
@@ -194,7 +194,12 @@ func TestPoolStatsPublisherRaisesHighDataUsage(t *testing.T) {
 		t.Fatalf("expected DataUsageHigh True at 85%%, got %+v", c)
 	}
 	select {
-	case <-rec.Events:
+	case e := <-rec.Events:
+		// The note is Eventf's format string, and this message carries usage
+		// percentages: it has to reach the operator verbatim.
+		if !strings.Contains(e, c.Message) {
+			t.Fatalf("the event must carry the condition message %q, got %q", c.Message, e)
+		}
 	default:
 		t.Fatal("expected a PoolUsageHigh event on first crossing")
 	}
diff --git a/internal/topology/conflict.go b/internal/topology/conflict.go
index 3cf3e5d..bb8dd65 100644
--- a/internal/topology/conflict.go
+++ b/internal/topology/conflict.go
@@ -22,6 +22,7 @@ package topology
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"slices"
 	"strings"
@@ -79,24 +80,28 @@ func (r *ConflictReconciler) Reconcile(ctx context.Context, _ ctrl.Request) (ctr
 		return ctrl.Result{}, err
 	}
 	topo := nodemap.FromNodes(list.Items)
+	// One node's failed update — a 409 against the agent's per-minute status
+	// writes, say — must not leave the rest of the fleet unstamped and so
+	// invisible to placement, so per-node errors are collected and joined.
+	// The requeue replays the whole pass: SetStatusCondition skips the nodes
+	// that did land, so the retry is cheap and emits no duplicate events.
+	var errs []error
 	for i := range list.Items {
-		// A mid-pass failure retries the whole pass: SetStatusCondition
-		// skips already-updated nodes, so the retry is cheap and emits no
-		// duplicate events.
 		if err := r.reconcileNode(ctx, &list.Items[i], topo); err != nil {
-			return ctrl.Result{}, err
+			errs = append(errs, fmt.Errorf("node %s: %w", list.Items[i].Name, err))
 		}
 	}
-	return ctrl.Result{}, nil
+	return ctrl.Result{}, errors.Join(errs...)
 }
 
 // reconcileNode updates one node's AddressConflict condition on change.
 func (r *ConflictReconciler) reconcileNode(ctx context.Context, node *miroirv1alpha1.MiroirNode, topo nodemap.Map) error {
 	cond := metav1.Condition{
-		Type:    ConditionAddressConflict,
-		Status:  metav1.ConditionFalse,
-		Reason:  reasonAddressUnique,
-		Message: "replication address is unique (or unset)",
+		Type:               ConditionAddressConflict,
+		Status:             metav1.ConditionFalse,
+		Reason:             reasonAddressUnique,
+		Message:            "replication address is unique (or unset)",
+		ObservedGeneration: node.Generation,
 	}
 	if topo[node.Name].AddressConflict {
 		// Group by the same canonical key the fold conflicts on: a raw
@@ -127,8 +132,9 @@ func (r *ConflictReconciler) reconcileNode(ctx context.Context, node *miroirv1al
 		return client.IgnoreNotFound(err)
 	}
 	if cond.Status == metav1.ConditionTrue && !wasConflicted && r.Recorder != nil {
+		// The message is data, not a format string: it embeds spec.address.
 		r.Recorder.Eventf(node, nil, corev1.EventTypeWarning, ConditionAddressConflict, "Reconcile",
-			cond.Message)
+			"%s", cond.Message)
 	}
 	return nil
 }
diff --git a/internal/topology/conflict_test.go b/internal/topology/conflict_test.go
index 1881595..c508e52 100644
--- a/internal/topology/conflict_test.go
+++ b/internal/topology/conflict_test.go
@@ -17,6 +17,8 @@ limitations under the License.
 package topology
 
 import (
+	"context"
+	"errors"
 	"strings"
 	"testing"
 
@@ -29,6 +31,7 @@ import (
 	ctrl "sigs.k8s.io/controller-runtime"
 	"sigs.k8s.io/controller-runtime/pkg/client"
 	"sigs.k8s.io/controller-runtime/pkg/client/fake"
+	"sigs.k8s.io/controller-runtime/pkg/client/interceptor"
 
 	miroirv1alpha1 "github.com/home-operations/miroir/api/v1alpha1"
 )
@@ -92,17 +95,55 @@ func TestConflictSinglePassCoversAllNodes(t *testing.T) {
 	}); err != nil {
 		t.Fatal(err)
 	}
-	for name, conflicted := range map[string]bool{nodeA: true, nodeB: true, nodeC: false} {
+	want := map[string]metav1.ConditionStatus{
+		nodeA: metav1.ConditionTrue,
+		nodeB: metav1.ConditionTrue,
+		nodeC: metav1.ConditionFalse,
+	}
+	// Errorf, not Fatalf: map iteration order is randomized, so aborting on
+	// the first unstamped node would report a different one each run and hide
+	// the rest — the blast radius is the whole point of this test.
+	for name, status := range want {
 		n := &miroirv1alpha1.MiroirNode{}
 		if err := c.Get(t.Context(), types.NamespacedName{Name: name}, n); err != nil {
 			t.Fatal(err)
 		}
-		cond := meta.FindStatusCondition(n.Status.Conditions, ConditionAddressConflict)
-		if cond == nil {
-			t.Fatalf("one pass must stamp the condition on %s", name)
+		if cond := meta.FindStatusCondition(n.Status.Conditions, ConditionAddressConflict); cond == nil ||
+			cond.Status != status {
+			t.Errorf("one pass must stamp AddressConflict=%s on %s, got %+v", status, name, cond)
 		}
-		if got := cond.Status == metav1.ConditionTrue; got != conflicted {
-			t.Fatalf("%s: conflicted=%v, want %v (%+v)", name, got, conflicted, cond)
+	}
+}
+
+// One node's failed status update must not leave the rest of the fleet
+// unstamped and so invisible to placement: the other nodes are still
+// stamped and the error surfaces for the requeue.
+func TestConflictOneBadNodeDoesNotBlockTheFleet(t *testing.T) {
+	base := newClient(t, addrNode(nodeA, "10.0.100.1"), addrNode(nodeB, "10.0.100.1"),
+		addrNode(nodeC, "10.0.100.3"))
+	// Stand in for the 409 the agent's per-minute status writes hand this
+	// reconciler mid-pass. The list is name-ordered, so node-a fails first.
+	c := interceptor.NewClient(base, interceptor.Funcs{
+		SubResourceUpdate: func(ctx context.Context, cl client.Client, sub string, obj client.Object, opts ...client.SubResourceUpdateOption) error {
+			if obj.GetName() == nodeA {
+				return errors.New("the object has been modified; please apply your changes to the latest version")
+			}
+			return cl.SubResource(sub).Update(ctx, obj, opts...)
+		},
+	})
+
+	r := &ConflictReconciler{Client: c}
+	_, err := r.Reconcile(t.Context(), ctrl.Request{NamespacedName: types.NamespacedName{Name: topologyRequestKey}})
+	if err == nil || !strings.Contains(err.Error(), nodeA) {
+		t.Fatalf("the per-node failure must surface with the node named, got %v", err)
+	}
+	for _, name := range []string{nodeB, nodeC} {
+		n := &miroirv1alpha1.MiroirNode{}
+		if err := c.Get(t.Context(), types.NamespacedName{Name: name}, n); err != nil {
+			t.Fatal(err)
+		}
+		if meta.FindStatusCondition(n.Status.Conditions, ConditionAddressConflict) == nil {
+			t.Errorf("%s must still be stamped after the failure on %s", name, nodeA)
 		}
 	}
 }
@@ -125,6 +166,19 @@ func TestConflictConditionRaisedAndNamesPeer(t *testing.T) {
 	}
 }
 
+// The condition records the generation it was computed from, so a True left
+// behind by a failed or not-yet-run pass is distinguishable from a current
+// one on a node whose address the operator just edited.
+func TestConflictConditionRecordsObservedGeneration(t *testing.T) {
+	a := addrNode(nodeA, "10.0.100.1")
+	a.Generation = 7
+	n := reconcile(t, newClient(t, a, addrNode(nodeB, "10.0.100.1")), nodeA)
+	if cond := meta.FindStatusCondition(n.Status.Conditions, ConditionAddressConflict); cond == nil ||
+		cond.ObservedGeneration != a.Generation {
+		t.Fatalf("expected observedGeneration %d, got %+v", a.Generation, cond)
+	}
+}
+
 // The fold conflicts on the parsed address; the peer list must group the
 // same way, or equal-but-differently-spelled IPv6 addresses raise a
 // condition naming no peer.
@@ -160,16 +214,27 @@ func TestConflictConditionClearsWhenResolved(t *testing.T) {
 }
 
 // The Warning event fires once per fresh conflict — repeat passes over an
-// unchanged (or message-only-changed) topology must stay silent.
+// unchanged topology, and passes that only rewrite an existing conflict's
+// peer list, must stay silent.
 func TestConflictEventFiresOncePerFreshConflict(t *testing.T) {
 	c := newClient(t, addrNode(nodeA, "10.0.100.1"), addrNode(nodeB, "10.0.100.1"))
 	rec := events.NewFakeRecorder(8)
 	r := &ConflictReconciler{Client: c, Recorder: rec}
-	req := ctrl.Request{NamespacedName: types.NamespacedName{Name: "topology"}}
+	req := ctrl.Request{NamespacedName: types.NamespacedName{Name: topologyRequestKey}}
 
 	if _, err := r.Reconcile(t.Context(), req); err != nil {
 		t.Fatal(err)
 	}
+	// A third node on the same address rewrites node-a's and node-b's peer
+	// list while their status stays True: the condition changes, so only the
+	// wasConflicted gate keeps them quiet. node-c is the one fresh conflict.
+	if err := c.Create(t.Context(), addrNode(nodeC, "10.0.100.1")); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := r.Reconcile(t.Context(), req); err != nil {
+		t.Fatal(err)
+	}
+	// Unchanged topology: no condition changes at all.
 	if _, err := r.Reconcile(t.Context(), req); err != nil {
 		t.Fatal(err)
 	}
@@ -179,7 +244,7 @@ func TestConflictEventFiresOncePerFreshConflict(t *testing.T) {
 	for range rec.Events {
 		fired++
 	}
-	if fired != 2 { // one per freshly conflicted node, none on the repeat
-		t.Fatalf("want exactly one event per fresh conflict (2), got %d", fired)
+	if fired != 3 { // one per freshly conflicted node, none on the repeats
+		t.Fatalf("want exactly one event per fresh conflict (3), got %d", fired)
 	}
 }
diff --git a/internal/topology/nodegroup.go b/internal/topology/nodegroup.go
index 45c7749..f4e3159 100644
--- a/internal/topology/nodegroup.go
+++ b/internal/topology/nodegroup.go
@@ -234,6 +234,9 @@ func (r *NodeGroupReconciler) SetupWithManager(mgr ctrl.Manager) error {
 		func(ctx context.Context, _ client.Object) []ctrl.Request {
 			groups := &miroirv1alpha1.MiroirNodeGroupList{}
 			if err := r.List(ctx, groups, client.UnsafeDisableDeepCopy); err != nil {
+				// A map func has nothing to requeue against, so the event is
+				// lost: log it rather than let the controller look idle.
+				ctrl.LoggerFrom(ctx).Error(err, "listing node groups to enqueue")
 				return nil
 			}
 			reqs := make([]ctrl.Request, 0, len(groups.Items))